Trustworthy Retrieval-Augmented Generation for the Digital Preservation and Transparent Presentation of Biomedical Scientific Heritage

Biomedical scientific heritage faces a dual challenge of exponential growth and inaccessibility. This paper presents a Trustworthy Retrieval Augmented Generation (RAG) framework designed for the digital preservation and transparent presentation of biomedical knowledge integrating ontology aware retrieval (MeSH/UMLS), hybrid BM25 + dense search via Reciprocal Rank Fusion, sentence level attribution, and a novel composite biomedical Trust Index (BTI) quantifying factual accuracy, attribution fidelity and adversarial robustness on the MedQuAD corpus, the framework achieves 89.1% top-5 recall, 85.6% attribution accuracy, F1 score of 75.2 and BTI of 0.84 on clean data.